We present a maximum likelihood estimation procedure for Shepard-Arabie ADCLUS (for “ADditive CLUStering”) model. The major advantages of our maximum likelihood ADCLUS (MLADCLUS) over theirs are two: (1) that MLADCLUS is capable of permitting hypothesis testing;(2) that MLADCLUS distinguishes and incorporates three types of models, “the representation model, ” “the error model, ” and “the response model, ” which are necessary for any reasonable psychological scaling. Applications of MLADCLUS to two different types of data gave the following results: (a) A solution with 13 semantic features such as “divide something three-dimensional into parts by force so as to cause damage, ” “cut something non-rigid into pieces by force, ” and “sever continuity” can give a good account of pair comparisons data of seven Japanese ‘break’ verbs.(b) An 8-cluster solution including such clusters as {P, R}, {F, P}, and {C, G} appears to best describe the structure of confusion matrix among six letters C, E, F, G, P, and R.(c) Both of these verbs and letters are better represented by the additive clustering model than by multidimensional scaling model.
View full abstract